""" 可视化消融实验：比较 SD-GSC, SAM-GSC, JEPA-GSC 的效果 类似于 vis_sd_featsv5.2.py，但同时展示三种方法的结果 """ import torch import torch.nn.functional as F import numpy as np import matplotlib.pyplot as plt from PIL import Image import os import sys from typing import Tuple # 添加项目路径 sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__)))) from refine_functions import ( refine_dino_with_sd, refine_dino_with_sam, refine_dino_with_ijepa, compute_dino_correlation, resize_attention ) # 图像预处理 class ImagePreprocessor: def __init__(self, size: int = 224): self.size = size self.mean = torch.tensor([0.485, 0.456, 0.406]).view(3, 1, 1) self.std = torch.tensor([0.229, 0.224, 0.225]).view(3, 1, 1) def __call__(self, img: Image.Image) -> torch.Tensor: img = img.convert("RGB") img = img.resize((self.size, self.size), Image.BILINEAR) img = torch.from_numpy(np.array(img)).float() / 255.0 img = img.permute(2, 0, 1) # HWC -> CHW img = (img - self.mean) / self.std return img.unsqueeze(0) def build_dinov2(device: str = "cuda"): """构建 DINOv2 模型""" hub_path = '/mnt/SSD8T/home/wjj/.cache/torch/hub/facebookresearch_dinov2_main' model = torch.hub.load(hub_path, 'dinov2_vitb14_reg', source='local').half() model = model.to(device).eval() for p in model.parameters(): p.requires_grad = False return model def extract_dino_features(model, image: torch.Tensor) -> torch.Tensor: """提取 DINO 特征""" with torch.no_grad(): features = model.get_intermediate_layers(image, reshape=True)[0] return features def visualize_similarity_comparison( image_path: str, dino_model, sd_attn: torch.Tensor, sam_attn: torch.Tensor, ijepa_attn: torch.Tensor, query_point: Tuple[int, int], refine_weight: float = 0.3, save_path: str = None, device: str = "cuda" ): """ 可视化比较三种方法的相似度图 Args: image_path: 输入图像路径 dino_model: DINOv2 模型 sd_attn: SD attention (B, HW, HW) sam_attn: SAM attention (B, HW, HW) ijepa_attn: I-JEPA attention (B, HW, HW) query_point: 查询点位置 (y, x) in patch coordinates refine_weight: refine 权重 save_path: 保存路径 device: 设备 """ # 加载和预处理图像 preprocess = ImagePreprocessor(size=224) image = Image.open(image_path) image_tensor = preprocess(image).to(device).half() # 提取 DINO 特征 dino_feats = extract_dino_features(dino_model, image_tensor) # (B, C, H, W) B, C, H, W = dino_feats.shape # 计算 DINO 相似度 dino_corr = compute_dino_correlation(dino_feats) # (B, HW, HW) # 调整 attention 尺寸以匹配 DINO target_size = H # 通常是 16 for 224x224 input with patch_size=14 if sd_attn is not None: sd_attn_resized = resize_attention(sd_attn, target_size) else: sd_attn_resized = None if sam_attn is not None: sam_attn_resized = resize_attention(sam_attn, target_size) else: sam_attn_resized = None if ijepa_attn is not None: ijepa_attn_resized = resize_attention(ijepa_attn, target_size) else: ijepa_attn_resized = None # Refine DINO 相似度 methods = { "Original DINO": dino_corr, } if sd_attn_resized is not None: methods["SD-GSC (Ours)"] = refine_dino_with_sd(dino_corr, sd_attn_resized, refine_weight) if sam_attn_resized is not None: methods["SAM-GSC"] = refine_dino_with_sam(dino_corr, sam_attn_resized, refine_weight) if ijepa_attn_resized is not None: methods["JEPA-GSC"] = refine_dino_with_ijepa(dino_corr, ijepa_attn_resized, refine_weight) # 获取查询点的相似度图 query_idx = query_point[0] * W + query_point[1] similarity_maps = {} for name, corr in methods.items(): sim_map = corr[0, query_idx].view(H, W).cpu().numpy() similarity_maps[name] = sim_map # 可视化 n_methods = len(similarity_maps) fig, axes = plt.subplots(1, n_methods + 1, figsize=(4 * (n_methods + 1), 4)) # 显示原图 axes[0].imshow(image.resize((224, 224))) axes[0].scatter([query_point[1] * (224 // W)], [query_point[0] * (224 // H)], c='red', s=100, marker='x') axes[0].set_title("Input Image") axes[0].axis('off') # 显示各方法的相似度图 for i, (name, sim_map) in enumerate(similarity_maps.items()): im = axes[i + 1].imshow(sim_map, cmap='hot', vmin=0, vmax=1) axes[i + 1].scatter([query_point[1]], [query_point[0]], c='cyan', s=50, marker='x') axes[i + 1].set_title(name) axes[i + 1].axis('off') plt.colorbar(im, ax=axes[-1], fraction=0.046, pad=0.04) plt.tight_layout() if save_path: plt.savefig(save_path, dpi=150, bbox_inches='tight') print(f"Saved visualization to {save_path}") plt.show() plt.close() def visualize_attention_comparison( sd_attn: torch.Tensor, sam_attn: torch.Tensor, ijepa_attn: torch.Tensor, query_point: Tuple[int, int], save_path: str = None ): """ 直接可视化三种方法的 attention map（不经过 DINO） """ H = W = int(sd_attn.shape[1] ** 0.5) if sd_attn is not None else int(sam_attn.shape[1] ** 0.5) query_idx = query_point[0] * W + query_point[1] attentions = {} if sd_attn is not None: attentions["SD Attention"] = sd_attn[0, query_idx].view(H, W).cpu().numpy() if sam_attn is not None: attentions["SAM Attention"] = sam_attn[0, query_idx].view(H, W).cpu().numpy() if ijepa_attn is not None: attentions["I-JEPA Attention"] = ijepa_attn[0, query_idx].view(H, W).cpu().numpy() n_attns = len(attentions) fig, axes = plt.subplots(1, n_attns, figsize=(4 * n_attns, 4)) if n_attns == 1: axes = [axes] for i, (name, attn_map) in enumerate(attentions.items()): im = axes[i].imshow(attn_map, cmap='viridis') axes[i].scatter([query_point[1]], [query_point[0]], c='red', s=50, marker='x') axes[i].set_title(name) axes[i].axis('off') plt.colorbar(im, ax=axes[i], fraction=0.046, pad=0.04) plt.tight_layout() if save_path: plt.savefig(save_path, dpi=150, bbox_inches='tight') print(f"Saved attention comparison to {save_path}") plt.show() plt.close() # ============ 主函数 ============ if __name__ == "__main__": import argparse parser = argparse.ArgumentParser() parser.add_argument("--image", type=str, required=True, help="Input image path") parser.add_argument("--query_y", type=int, default=8, help="Query point y (in patch coords)") parser.add_argument("--query_x", type=int, default=8, help="Query point x (in patch coords)") parser.add_argument("--refine_weight", type=float, default=0.3) parser.add_argument("--save_dir", type=str, default="./ablation_results") args = parser.parse_args() os.makedirs(args.save_dir, exist_ok=True) device = "cuda" if torch.cuda.is_available() else "cpu" print("Building DINOv2 model...") dino_model = build_dinov2(device) # TODO: 加载预提取的 attention 或实时提取 # 这里用随机 attention 作为示例 print("Using dummy attention for demonstration...") HW = 256 # 16x16 patches sd_attn = F.softmax(torch.randn(1, HW, HW, device=device), dim=-1) sam_attn = F.softmax(torch.randn(1, HW, HW, device=device), dim=-1) ijepa_attn = F.softmax(torch.randn(1, HW, HW, device=device), dim=-1) query_point = (args.query_y, args.query_x) print("Visualizing similarity comparison...") save_path = os.path.join(args.save_dir, "similarity_comparison.png") visualize_similarity_comparison( args.image, dino_model, sd_attn, sam_attn, ijepa_attn, query_point, args.refine_weight, save_path, device ) print("Done!")